Road detection via entropy By Anna Zaidman 1 1 What is entropy? - PowerPoint PPT Presentation

Road detection via entropy By Anna Zaidman 1 1

What is entropy? Entropy is a mathematically - defined thermodynamic quantity that helps to account for the flow of energy through a thermodynamic process. 2

What is entropy? As a mathematical function, entropy is used to measure the level of “disorder” within a certain sample of values, a certain entropy value is assigned to a pixel by measuring entropy of a sample of pixel values present in a given “window” around that pixel. Smooth and visually “uniform” parts of the image thus have low or very low entropy (no matter what are the real pixel values - “color”), while areas with higher diversity of pixel values, especially with sudden shifts in image color – gain higher entropy values.

What is entropy? If we assume that the road to follow has a consistent color or texture, we can state that the neighborhoods of pixels forming that road have lower diversification than a neighborhood of pixels forming people, trees, cars, traffic signs, and so on, which might be composed of different levels of illumination, textures, patterns and edges. Lack of diversity means also, lack of information, when the opposite implies lot of information. 4

My goal is to detect a drivable road using an adaptive entropy filter .

How to create an entropy filter? Let's define the entropy H of an image I , as the amount of information that exists over a statistically distributed set of pixels 𝑕 , over a window w , where for each grayscale level i its probability of appearance is 𝒒(𝒉 𝒋 ) , as shown in equation (1). The probability of a graylevel i is calculated according to the histogram of the image as shown in equation (2). (1) 𝐼(𝑕) = − 𝑞 𝑕 𝑗 𝑚𝑝𝑕 (𝑞 𝑕 𝑗 ) 𝑥 ℎ𝑗𝑡𝑢(𝑕 𝑗 ) (2 ) 𝑞 𝑕 𝑗 = 𝑓𝑜𝑒 ℎ𝑗𝑡𝑢(𝑕 𝑘 ) 𝑘=1 Thus , an entropy filter consists of calculating the entropy over all the pixels in the image, using a predefined mask of neighborhood. 6

7

8

9

Implementation 10

Conclusions The entropy filter turned out to be a robust technique to perform segmentation and Clustering. A setback is the implementation of this method in real time. Due to the nature of the entropy filter , the computational load is very expensive. 11

How can we resolve it? 1. In order to improve performance , look-up-tables with a priori data could be used to compute the entropy itself . 2. Use other filters with similar results but faster computing. 12